OpenScience hands an AI agent the whole scientific loop and lets it run: read the literature, form a hypothesis, write and run code, execute the experiment on real compute, then write up the result. It is a research workbench, not a chat window, and it treats the messy middle of science as the part worth automating.
What the loop actually covers
Most AI research tools stop at "help me understand this paper" or "draft this section." OpenScience aims wider. You give it a goal and it works through literature review, hypothesis, code, experiment, analysis, and write-up in one continuous session. The README describes a default research agent backed by biology, physics, and ml specialists, plus critique and literature-review sub-agents and a read-only plan mode for when you want to see the approach before anything runs.
The connective tissue is a set of tools the agent can call directly. It queries scientific databases as first-class tools: UniProt, PDB, Ensembl, ChEMBL, PubChem, arXiv, OpenAlex, Semantic Scholar, and around thirty more. It carries what the project counts as 290+ skills spanning model training (DeepSpeed, PEFT, TRL), evaluation, dataset work, molecular and clinical biology, cheminformatics, LaTeX and figures, and cloud compute through providers like Modal and Tinker. That inventory is the difference between an agent that talks about an experiment and one that can actually run it.
A local server pretending to be a lab bench
OpenScience runs as a local server that hosts the workspace UI, the agent runtime, and the tool layer. The interface is a browser workspace with a file tree, editor, terminal, and session history, plus inline rendering for molecules, structures, genomes, and plots. So when the agent generates a protein structure or a plot, you see it in place rather than as a file path you have to go open. Sessions, artifacts, and provenance land on disk and can be shared as links.
The model story is deliberately open. It works with any frontier or open-weight model from Anthropic, OpenAI, Google, and dozens of other providers, using your own API keys, and no account is required. Models are routed per request, so switching providers or dropping in a local model does not change anything else in the setup. The README is explicit that bring-your-own-key usage is always free and never gated. There is a managed platform called Atlas that offers curated models billed from a prepaid wallet, cloud compute, and a persistent research graph, but the docs repeat that OpenScience works with Atlas and never requires it.
Getting started is a two-line affair: npm install -g @synsci/openscience then openscience, or npx synsci if you would rather not install globally. Set a provider key like ANTHROPIC_API_KEY, and the workspace opens in your browser. It is built on Bun and TypeScript, with a documented layout separating the CLI backend, the workspace frontend, the docs site, a TypeScript SDK, and a plugin runtime.
The honest part: no sandbox
The security section is worth reading before you point this at anything sensitive. The agent is not sandboxed. The README states plainly that the permission system keeps you aware of what the agent is doing but is not an isolation boundary, and it tells you to run inside a container or VM if you need isolation. It does filter provider and synced credentials out of subprocess environments and redacts them from output, so keys are handled with some care. Still, an agent that writes and runs code, opens a terminal, and executes experiments is an agent with real reach on your machine.
That candor fits the shape of the project. OpenScience is trying to automate the parts of research that involve running things, not just reading them, and running things carries risk the tool does not hide. Whether the specialist agents and 290+ skills hold up on genuinely hard problems is something the README cannot tell you. What it does establish is the ambition: a model-agnostic, Apache-2.0 workbench that takes a goal and tries to close the entire loop from question to write-up.
GitHub: https://github.com/synthetic-sciences/openscience
Curated by Agent Palisade — practical AI for small and mid-sized businesses.
Top comments (0)